Mobile Netw Appl
DOI 10.1007/s11036-010-0221-2
An Energy Efficient Middleware Architecture
for Processing Spatial Alarms on Mobile Clients
Anand Murugappan · Ling Liu
© Springer Science+Business Media, LLC 2010
Abstract Time based alarms are used by many on a
daily basis. Spatial alarms extend the very same idea to
location based triggers, which are fired whenever a mobile user enters the spatial region of the location alarms.
Spatial alarms provide critical capabilities for many
mobile location based applications ranging from personal assistants, inventory tracking to industrial safety
warning systems. In this paper we present a middleware
architecture for energy efficient processing of spatial
alarms on mobile clients, while maintaining low computation and storage costs. Our approach to spatial
alarms provides two systematic methods for minimizing
energy consumption on mobile clients. First, we introduce the concept of safe distance to reduce the number
of unnecessary mobile client wakeups for spatial alarm
evaluation, enabling mobile clients to sleep for longer
intervals of time in the presence of active spatial alarms.
We show that our safe distance techniques can significantly minimize the energy consumption on mobile
clients compared to periodic wakeups while preserving
the accuracy and timeliness of spatial alarms. Second,
we develop a suite of techniques for minimizing the
number of location triggers to be checked for spatial
alarm evaluation upon each wakeup. This further reduces the computation cost and energy expenditure on
mobile clients. We evaluate the scalability and energyefficiency of our approach using a road network simulator. Our spatial alarms middleware architecture offers
A. Murugappan (B) · L. Liu
College of Computing, Georgia Institute of Technology,
Atlanta, GA, USA
e-mail: anandm@cc.gatech.edu
L. Liu
e-mail: lingliu@cc.gatech.edu
significant improvements on battery lifetime of mobile clients, while maintaining high quality of spatial
alarm services, especially compared to the conventional
approach of periodic wakeup and checking all alarms
upon wakeup.
Keywords spatial alarms · location based alarms ·
location based services · mobile applications ·
energy efficient algorithms
1 Introduction
Spatial alarms are considered by many as one of the
critical mobile location-based applications in future
computing environments. For instance, a user could set
a spatial alarm on her mobile client (e.g., her smart
phone or her personalized navigation system in her
car), which alerts her whenever she is nearby the dry
clean store or the grocery stores in her neighborhood,
reminding her to pick up or drop off her dry cleaning items or automatically popping up her grocery
shopping list as a spatial reminder. Furthermore, the
mobile user can install a spatial alarm targeted at a
particular location (e.g., her favorite dry clean store)
with a specific music piece or with a voice recording
or with an image of the store. Upon her arrival of the
alarm region, the music piece or the audio and video
will play on her mobile client (e.g., cellphone). One
popular way to install such personalized spatial alarms
is to use the map software on the mobile client of the
user by marking the area of interest as the location
alarm area, with associated actions upon the firing of
corresponding location triggers. Most of the existing
research on spatial alarm type of mobile or wireless
Mobile Netw Appl
applications are produced from the Human–Computer
Interaction community, which focuses on the usability
study of location reminders on mobile clients or the
efficient usage of multimedia to register and notify of
location reminders on a mobile client on behalf of the
human user of the device. Surprisingly, there are little
research to date that is dedicated to energy efficient
evaluation of spatial alarms on mobile clients.
Processing of spatial alarms requires meeting two
demanding objectives: high accuracy, which ensures no
alarms are missed and high energy efficiency and high
scalability, which not only minimizes the unnecessary
processing cost and the consumption of energy on spatial alarm processing but also scales the energy efficient
processing to larger number of spatial alarms on mobile clients. The conventional approach to designing
the middleware architecture for spatial alarms involves
periodic alarm checks at a high frequency. Each spatial
alarm check is conducted by testing whether the user is
entering the spatial region of the alarm. High frequency
is essential to ensure that none of the alarms are missed.
This technique is simple but can be extremely energy
inefficient due to both frequent wakeups and evaluation of all alarms upon each wakeup. This is especially
true when the mobile client is traveling in a location
that is distant from the spatial areas of all her location
triggers or when the collection of spatial alarms is set
on spatial regions that are far apart from one another.
In addition, some types of mobile clients have stronger
resource constraints such as smart phones and handheld PDAs compared to navigation systems in cars.
In this paper we present our architecture for energy
efficient processing of spatial alarms on mobile clients,
while maintaining low computation and storage costs.
We present two systematic methods that can progressively minimize the amount of energy consumption on
mobile clients for all types of spatial alarms. The first
method utilizes the concept of safe distance to reduce
the number of unnecessary wakeups on mobile clients
for spatial alarm evaluation. By enabling mobile clients
to sleep for longer intervals of time in the presence
of active spatial alarms, we show that our safe distance techniques can significantly minimize the energy
consumption on mobile clients compared to periodic
wakeups, while preserving the accuracy and timeliness
of spatial alarms. The second mechanism focuses on
alarm checks upon each wakeup. We develop a suite
of techniques for minimizing the number of location
triggers to be checked upon each wakeup for different
types of spatial alarms. This allows us to further reduce
the computation cost and energy expenditure on mobile clients. Our experimental evaluation using a road
network simulator shows that our spatial alarms mid-
dleware architecture offers significant improvements
on battery lifetime of mobile clients, while maintaining
high quality of spatial alarm services compared to the
conventional approach of periodic wakeup and checking all alarms upon a wakeup.
2 System overview
A spatial alarm consists of three components: the spatial region on a two-dimensional geographical plane,
the action to be taken upon firing of the alarm, and
the alarm termination condition, usually a temporal
event such as time point or time interval. The spatial
regions used in spatial alarms can be of any shape.
We capture each of such spatial regions by a rectangular bounding box, denoted by (x1 , y1 , x2 , y2 ), where
(x1 , y1 ) and (x2 , y2 ) represent the top-left and bottomright vertices of the bounding rectangle. Without loss
of generality, in the rest of the paper, we simply assume
that each mobile client can install n spatial alarms (n ≥
0) and all spatial alarms are expressed by a rectangle
spatial region, denoted by Ai for 1 ≤ i ≤ n. In our first
spatial alarm client middleware prototype, we use a
system supplied default spatial range in the absence
of spatial region specification of a user-defined alarm.
Each mobile client can install as many spatial alarms as
the user wishes over the geographical area of interest.
Multiple mobile clients can set spatial alarms on the
same locations.
In this paper we assume that mobile clients have
limited energy, storage, and computational resources.
We also assume that at least one of positioning technologies, such as GPS, WiFi based triangulation or cellular network assisted location identification services,
is available for the mobile client to identify its current
location. Each mobile client is a moving object with an
accompanying mobile device, such as cell phone, PDA,
car, which computes and communicates the current location of the mobile client with the Spatial Alarm client
middleware. In addition, mobile clients may have a
map service available on their portable device, through
which the mobile client can navigate on the map to
identify and install their personalized spatial alarms
through interactive mode of communication with the
Spatial Alarm client middleware. Batch installation of
spatial alarms is also possible.
Spatial alarms differ from spatial location queries in
a number of ways. First, spatial queries such as “tell
me the gas stations within 10 miles on the highway
85 north” require continuous evaluation of the queries
as the mobile client moves on the highway 85 north.
However, spatial alarms, such as “notify me whenever I
Mobile Netw Appl
am 5 miles away from this particular dry cleaning store
(marked on the map)”, only requires the alarm to be
evaluated when the mobile client moves to a region
that is within 5 miles of the specific dry cleaning store.
Thus, it is no use to wake up a mobile client if she
is 30 miles away from the dry cleaning store. Clearly,
the movement patterns of the mobile client and the
distance from the current location of a mobile client
to all her alarms are the two critical factors that will
affect when the mobile client needs to wakeup and
what alarms need to be checked upon each wakeup.
Thus one can optimize the spatial alarm processing by
devising more energy efficient algorithms.
Mobile devices conserve energy by spending most of
their time in a low energy state such as sleep mode.
Hence one of the critical design objectives for client
middleware architecture is to minimize the number of
device wakeups in spatial alarm processing. For instance, the 206 MHz Itsy [3] pocket computer spends
540 mW power in the System Idle, 0% processor idle
state, spends 100 mW power in the System Idle, 95%
processor idle state, and while in the Sleep mode it
just spends 8.39 mW power (which is about 64 times
lesser than the 0% processor idle case). It is interesting
to note that mobile devices like in the case of Itsy
computer [3] have a battery lifetime of only 3.8 h when
running in the high energy System Idle, 95% processor
idle state, while in the Sleep mode the lifetime is as high
as 279 h.
The conventional approach for implementing a location based reminder system is to wake up the device and
check the alarm conditions periodically. If the period
is too large the mobile device might miss alarms since
there may be situations where the mobile client passes
through the ‘alarm area’ while asleep (between periodic
checks). Hence, to reduce the number of alarm misses,
the wakeup period would have to be kept small enough.
The smallest wakeup period can be set using the location update frequency (e.g., GPS sampling period).
Clearly, the periodic check approach would be very
energy inefficient. Also it is important to note that if
the mobile client is far away from any of her alarms
then depending on the maximum speed of the client, it
is possible to sleep for longer durations of time and still
guarantee that none of the alarms would be missed.
In the context of spatial alarm processing, one can
save energy by two phase optimizations. In Phase one
we minimize the number of device wakeups and in
phase two we further minimize the number of alarms
checked upon each wakeup. The phase one optimization helps keep the mobile client in the ‘Sleep mode’ as
long as possible, while the phase two optimization helps
reduce the computation cycles used for alarm evalua-
tion, which further reduces the energy requirements. In
addition to energy efficiency, another important goal of
spatial alarm processing is to maintain the low or zero
alarm misses.
In this paper we propose the concept of safe period
which computes the time period during which the mobile client can continue to sleep without missing any
of her alarms. We compute this safe period using the
distance from the current location of the mobile client
to all her alarms and the speed measure of the mobile
client. We consider Euclidean distance and road network distance as two alternative distance functions and
consider maximum travel speed and expected travel
speed of a mobile client as two alternative speed functions. The expected speed measure is used to handle the
situation where the mobile clients do not travel at their
maximum possible speed at all times, hence by considering the ‘Average Speed’ or the ‘Expected Speed’, we
may present a wakeup algorithm that is more adaptive
to the movement behavior and the actual distance from
the client to the alarms. Further, by considering road
network distance as a more accurate prediction of the
distance from the mobile client to all her alarms, we can
further extend the sleep time without any alarm misses.
We discuss these alternative approaches to minimizing
alarm wakeups in greater detail in Section 3.
In addition to minimizing wakeups in the phase one
optimization, we have noted that it is also essential
to reduce the computations performed for processing
alarms at each wakeup. The naive approach requires
checking against all alarms upon each wakeup, which
can be expensive. One way to mitigate this problem is
to group the alarms based on their spatial proximity,
check them together in groups and drill down until
individual alarm checks are performed. Thus our phase
II optimization focus on minimizing alarm checks with
three strategies. We extend three popular indexing algorithms to perform the alarm grouping optimizations
using R-Tree [13], Voronoi diagram that partitions
the two-dimensional coordinate space into Voronoi
Regions [2, 11], resulting in a very efficient nearest
alarm lookup. To mitigate the additional storage cost,
we use the road network information to generate a
Network Voronoi Diagram [8], which achieves very
similar performance numbers as the Euclidean distance
metrics, but with reduced storage costs. We discuss
these optimizations in greater detail in Section 4.
3 Phase I optimization: minimizing device wakeups
Our middleware architecture for spatial alarm processing consists of two phase optimizations. In this section
Mobile Netw Appl
we discuss the phase one optimization strategies that
minimize the number of device wakeups. Apart from
the high energy consumption, an important problem
with the periodic wakeup approach is that it is hard to
estimate how frequently the device should wakeup to
ensure no alarms will be missed. Two factors that are
critical in determining such a frequency: (a) The speed
of the mobile client; and (b) the size of the spatial alarm
region. Unless the frequency is set to be extremely
high (close to the location update frequency), it would
always be possible to introduce cases where alarms
can be missed by having alarms of the size smaller
than the distance traveled by the mobile client between
two consecutive wakeups. Thus the key challenge is to
determine the right time for mobile clients to wakeup
in terms of energy efficiency and alarm accuracy and
given a location update, how to determine the subset of
alarms that should be checked to conserve energy while
maintaining zero alarm misses.
With both the problem of guaranteed alarm delivery
and that of energy conservation in mind, we propose
four optimization strategies to estimate the safe period
and use aperiodic wakeup of the mobile client based on
the distance of the client to all her alarms and the travel
speed of the mobile client. The rest of this section is organized as follows. First, we describe two different techniques for estimating the distance the mobile client’s
current location to any of her alarms. Second, we describe the maximum speed and the expected speed as
two alternative measures to the speed of mobile client.
Finally, we present the four optimization algorithms
that estimate the time to sleep, or the so-called the safe
period, for the mobile client before her next wakeup.
Our approach also ensures that alarms can be added
and removed while the system is operational and such
dynamic alarm addition and deletion do not interfere
with the safe period estimation of the mobile client.
the original map into tiles and selectively download
relevant tiles to a mobile client.
3.1.1 Euclidean distance to an alarm
Given a spatial alarm Ai with rectangular spatial alarm
region represented by four vertices of the rectangle: (P1 , P2 , P3 , P4 ) where P1 = (x1 , y1 ), P2 = (x2 , y1 ),
P3 = (x2 , y2 ) and P4 = (x1 , y2 ). Let the mobile client
be at Pm represented by the coordinates (xm , ym ), then
the Euclidean distance from Pm to the alarm region of
Ai , denoted by d Ai , can be computed by considering
four cases. Case 1: when the mobile device is within
the alarm boundaries the distance to the alarm is zero;
Case 2: when the mobile device is within the y scope
(represented using dotted lines in Fig. 1a) the distance
is the shortest of the distances to alarm edges parallel
to the x axis from the mobile client; Case 3: when
the mobile device is within the x scope the distance
is the shortest of the distances to alarm edges parallel
to the y axis from the mobile client; and Case 4: when
the mobile device is outside both the x and y scope then
the distance is the minimum of the Euclidean distances
to the four vertices. The four cases can be formally
defined as follows:
d Ai
⎧
0
⎪
⎪
⎪
⎪
⎨
= min(|xm − x1 |, |xm − x2 |)
⎪
⎪
⎪
min(|ym − y1 |, |ym − y2 |)
⎪
⎩
min(Dm1 , Dm2 , Dm3 , Dm4 )
x1 ≤ xm ≤ x2
and y1 ≤ ym ≤ y2
y1 ≤ ym ≤ y2 only
x1 ≤ xm ≤ x2 only
otherwise
Where Dm1 , Dm2 , Dm3 , Dm4 denote the Euclidean
distance from Pm to the four rectangle vertices
P1 , P2 , P3 , P4 respectively. The distance function Dij =
(xi − x j)2 + (yi − y j)2 is used to compute the Euclidean distance between two points Pi and P j.
3.1.2 Road network distance
3.1 Measuring the distance to alarm
There are two most commonly used methods for measuring the distance from a mobile client’s current location to an alarm. They are (a) Euclidean Distance and
(b) Road Network Distance. The Euclidean Distance
approach is simpler and requires much lesser data but
may at times underestimate the time to sleep before
the next wakeup. The Road Network Distance measure
offers a more accurate estimate of the distance from a
mobile client’s current location to the spatial region of
the alarm, but it introduces additional overhead with
handling the road network map data. We propose techniques to mitigate this additional overhead by dividing
One of the main weaknesses of the Euclidean Distance
measure is that the estimated distance is often shorter
than the actual distance that the mobile client would
have to travel to get to the spatial region of interest
of a given alarm due to the underlying traversal
restrictions imposed by the road network. The Road
Network Distance measure uses the Dijkstra’s shortest
path algorithm [5, 7] to estimate the distance from
the mobile client’s current location to an alarm as
shown in Fig. 1b. The underlying road network is
represented by the solid line and the mobile client is
represented by a shaded circle labeled by 1. Since the
mobile client is restricted to move along the roads,
Mobile Netw Appl
(a) Euclidean Alarm Distance
(b) Road Network Alarm Distance
(c) Legend
Fig. 1 Distance to alarm from mobile client
the only places where it can enter the alarm area
would be the points of intersection of the alarm with
the roads, denoted by S1 , S2 , S3 in Fig. 1b. Hence, in
order to estimate the Road Network Distance (Rd Ai )
from Mobile Client at Pm to an alarm Ai we calculate
the shortest network distance (using Dijkstra’s
algorithm) to each of the three points of intersection
and choose the minimum. Mathematically, Rd Ai =
min(SPath(Pm , S1 ), . . . , SPath(Pm , S j), . . . , SPath(Pm ,
Sk )) where k is the total number of intersections of
the alarm area with different roads on the map, S j
represents the jth point of intersection (1 ≤ j ≤ k), and
SPath(Pm , S j) representing the Shortest Road Distance
from Mobile Client’s location at Point Pm to the jth
alarm-road intersection point S j obtained using the
Dijkstra’s Shortest path Algorithm [5, 7].
Since computing the road network distance requires
detailed maps to perform the calculations, we need
efficient mechanisms to handle the situation when it is
not possible to store the entire available map information on the mobile client. One approach to mitigate this
problem is to divide the entire map into square map
tiles and fetch only the relevant tiles to the memory
of the mobile client each time when the road network
distance is computed. If the mobile client does not have
memory to host the map for the entire geographical
area of interest, a third party map service can be used
for mobile clients to fetch the relevant titles from the
server. Consider the example shown in the Fig. 2. Given
the midtown Atlanta map as the area of interest as
shown in Fig. 2a. To begin with, the mobile client
requests the server to send over nine equal sized square
tiles that form a 3 by 3 tile matrix on the map with the
tile in which the mobile client resides as the inner most
tile (the tile numbered 13), and eight tiles surrounding
the inner most tile are numbered 7, 8, 9, 12, 14, 17, 18
and 19, as shown in Fig. 2b. The mobile client then creates an internal system def ined alarm over the tile numbered 13 such that whenever the mobile client moves
outside the tile 13, an additional tile fetch request will
be issued to the map server. For instance, Several factors affect the decision on the tile size. On one hand, the
larger the tile size is, the higher the energy conservation
will be since the client will be further away from the
internal system alarm on the current tile. On the other
hand, the larger the tile size is, the more demand will
be set on the client storage capacity. Thus a proper
setting of tile size needs to trade off between the storage
constraint and the energy conservation need. The good
news is that even assuming low storage availability of
Fig. 2 Mobile client map window: tiling of maps to reduce storage costs
Mobile Netw Appl
2 MB, detailed road map [1] of a medium size city, such
as Atlanta or Washington DC, can be stored without
having to request additional tiles from the map server.
In the rest of this paper we assume that the available
storage is limited and hence the tile size is small. We
consider the Georgia Institute of Technology campus in
Atlanta (which is about 1.61km2 ) to be contained in its
entirety in a tile as shown in Fig. 2.
When α = 1 this function would return the EWMA
speed. In the experiments we set α = 0.5 and β = 0.8.
p
vexpected = 0
c
vexpected
=β∗
(1)
D(lc , l p )
p
+ (1 − β) ∗ vexpected
tc − t p
c
vexpected = α ∗ vexpected
+ (1 − α) ∗ vmax
3.2 Speed estimation methods
Maximum Speed: The use of maximum travel speed
of the mobile client has a number of advantages and
disadvantages. On the brighter side, one can set the
‘Maximum travel speed’ by pre-configuration based on
either the nature of the mobile client (such as a car
on the move or a pedestrian walking on the street),
or depending on the types of roads used, or the environmental limitations on the mobile client such as
entered a building. For instance, if the mobile client
is always expected to be attached to a car then the
maximum speed of the car or the highest speed limit
on the road on which the car travels can be used as the
‘Maximum travel speed’. However, using the maximum
travel speed as the speed estimation technique is often
over pessimistic since it cannot adapt to the situation
where a mobile client may stop for an extended period
of time or may suddenly turn onto a road with very low
speed limit.
Expected Speed: If in most cases the world is more
benign, then the speed of the mobile client will often
be lesser than the maximum speeds. This is especially
true for road networks that have different traffic patterns at different times of a day, or that have winding
roads or other unexpected conditions. Hence, a more
pragmatic approach is to estimate the expected speed of
the mobile client and use the expected speed measure
to determine the safe period (the time to sleep) before
the mobile client needs to wake up the next time. In
the first spatial alarm middleware prototype system, we
calculate the current expected speed using the Exponential Weighted Moving Average (EWMA) scheme
(Eq. 9) based on current and previous location of the
mobile client and the previous expected speed estimate.
In addition to the current expected speed weighted
by α, the future expected speed is increased by performing a weighted average (1 − α) with the maximum
speed to ensure higher guarantee of alarm delivery
and zero or lower alarm misses (Eq. 10). The lower
the α value is, the better the alarm delivery guarantee
will be, but the higher the incurred energy cost. When
α = 0 this function would return the Maximum Speed.
(2)
(3)
p
c
where vexpected , vexpected
, vexpected are the previous, the
current, and the future expected travel speed of the mobile client respectively, tc and t p represent the current
and previous time instances, and lc and l p represent the
current and the previous location of the mobile client at
time instances tc and t p respectively.
Note that client travel speed estimation is essential
in determining the number of wakeups to be used. Our
speed estimation formula does not assume that the mobile client knows its exact travel speed. Instead, we use
the current and previous location of the mobile client,
her current and previous wakeup time instances, and
the previous expected speed to estimate her current and
her future expected travel speeds. By taking into consideration the maximum velocity, the current and the
past movement behavior, we show that our expected
speed calculation presents a reasonable approximation
to the future travel speed of the mobile client.
3.3 The four aperiodic wake-up algorithms
In this section, we introduce the concept of safe period
based on the distance function and the speed function,
and present four safe period based wakeup algorithms
by combining the distance functions with the speed
functions.
Tsleep = min(d A1 ...d Ai ...d An )/vmax
(4)
Tsleep = min(d A1 ...d Ai ...d An )/vexpected
(5)
Tsleep = min(Rd A1 ...Rd Ai ...Rd An )/vmax
(6)
Tsleep = min(Rd A1 ...Rd Ai ...Rd An )/vexpected
(7)
where Tsleep is time duration for which the mobile
client can sleep without potentially missing delivery of
any alarm, n is the total number of alarms installed
on the mobile client, d Ai , Rd Ai are the euclidean and
road network distances from the mobile client’s current
location to the ith spatial alarm Ai (1 ≤ i ≤ n) and vmax ,
Mobile Netw Appl
vexpected are the maximum and expected travel speeds of
the mobile client as defined below.
p
vexpected = 0
c
=β∗
vexpected
(8)
D(lc , l p )
p
+ (1 − β) ∗ vexpected
tc − t p
c
vexpected = α ∗ vexpected
+ (1 − α) ∗ vmax
p
(9)
(10)
c
where vexpected , vexpected
, vexpected are the previous, the
current, and the future expected travel speed of the
mobile client respectively, tc and t p represent the current and previous time instances, lc and l p represent the
current and the previous location of the mobile client
at time instances tc and t p respectively and D is the
distance function.
The parameter α is a weigh factor that is used to
compute the expected velocity of a mobile in the near
future by balancing between the current velocity of the
mobile and the past maximum travel speed of the same
mobile client. The value alpha varies between 0 and 1.
If α is set close to 1 then the expected future velocity
is computed primarily based on the current velocity of
the mobile. If α is set close to 0, then the expected
future velocity is computed primarily based on the past
maximum-speed of the mobile.
We choose a value that is right in the middle (i.e.,
0.5) between the two points. The objective of this α
setting is to compute the future expected velocity based
on equal weight of both the current velocity and the
past maximum velocity of the client. This allows us
to predict the future expected velocity of a mobile by
combining her current velocity and her maximum travel
speed.
β is also a parameter with value between 0 and 1. It is
used as a knob to control how quickly the system adapts
to changing speeds. A higher value would result in the
system responding faster to changes in speeds while
a lower value would result in slower adaptation while
placing more weight on the historical values. We chose
0.8 to keep the adaptability relatively high but slightly
lower than 1 to be influenced by historical values as
well.
Safe Distance with Max Speed defines the safe distance of a mobile client to each of her spatial alarms
by combining the Euclidean Distance function and the
maximum speed as shown in Eq. 4. At each wakeup,
the mobile client will perform two tasks—(a) process
the spatial alarms installed on the client (b) estimate
the safe period (i.e. the time to sleep) denoted by Tsleep ,
before the next wakeup. Safe Distance with Expected
Speed in Eq. 5 exploits the fact that not all mobile
clients might travel at the maximum possible speed
at all times. Since ’expected speed’ would ideally be
lesser than the maximum speed, the number of device
wakeups can be further reduced at the expense of potentially missing some alarms. Safe Road Distance with
Max Speed in Eq. 6 exploits the limitations imposed on
the movement of the mobile client by the underlying
road network while ensuring no alarms are missed.
Safe Road Distance with Expected Speed in Eq. 7 is
a natural extension and combines the advantages of
the ’expected speed’ approach with the ’road distance’
approach.
Figure 3 provides an example scenario to illustrate
the road distance based wakeup algorithms with the
naive period wakeup approach. Circles in all three cases
represent wakeups of the mobile client and rectangles
represent the spatial alarms installed on the mobile
client. The spatial alarm layout on the road network is
shown in all three scenarios. Figure 3a represents the
periodic Wake-Up strategy. The shaded path consists
of many circles one overlapped with another, showing the high frequency of the wakeups in this case.
Figure 3b shows the case of Safe Distance with Max
Speed. It clearly demonstrates the significant reduction
in the number of wakeups comparing to the Periodic
wakeup case. Figure 3c represents the case of Safe
Road Distance with Expected Speed. It shows a further
reduction on the frequency of the wakeups compared to
the safe road distance with maximum speed in Fig. 3b.
It is interesting to note that some of the wakeups are
slightly further apart when the mobile client travels
away from the spatial alarms in terms of road network
distance. This example also illustrates the contrast between Euclidean distance and road network discussion.
For instance, the spatial alarms on the lower left bottom
of the map, though close to the mobile client on the
move in terms of Euclidean space, are relatively further
away in terms of actual road network distance. Hence
by using Safe Road Distance, a considerable number of
unnecessary wakeups can be avoided.
4 Phase II optimization: minimizing alarms checked
The conventional approach is to periodically wakeup
the mobile client and to check against all the alarms
individually upon each wakeup. We have shown that
periodic wakeup is naive and consumes a lot of energy
unnecessarily. We have shown in the previous section
that by utilizing the distance to alarms and the estimate
of travel speed of the mobile client, our safe period
based wakeup algorithms significantly outperform the
periodic wakeup approach in terms of energy efficiency
Mobile Netw Appl
(a) Periodic
(b) SDMax Speed
(c) Road SD Exp Speed
Fig. 3 Comparison of wake up strategies
without introducing loss of alarms. In this section we
argue that the approach of check all alarms upon each
wakeup is naive and waste of resources. We describe
three approaches to minimize the number of alarm
checks per wakeup and show how these approaches can
reduce the computation cost and the energy consumption involved in alarm checks.
In the first approach, we group spatial alarms that
are in close spatial proximity in a hierarchical fashion.
Alarm Checking happens in groups, thus minimizing
the overall number of alarm checks performed per
wakeup. In the second approach, we divide the geographical area of interest into Voronoi regions based
on Euclidean distance to the alarms with each region
storing information that can quickly identify the nearest alarm in the vicinity. Upon each wakeup, alarm
checks are performed only against the ’nearest’ alarm
by looking up the information in the Voronoi region in
which the mobile client currently resides. In the third
approach, we further reduce the storage cost involved
with the Euclidean distance based Voronoi diagram
approach by dividing the road network map in the geographical area of interest into network distance based
Voronoi regions, based on the number of road network
nodes, using Network Voronoi Diagrams. We below
describe each of these three alarm check optimizations
in detail.
4.1 Hierarchical grouping of alarms
When the geographical area in which a mobile client
installs her alarms is big, the number of alarms installed
is large and distributed across the entire area of interest,
checking all alarms upon each wakeup is not only unnecessary but also a clear waste of resources. We first
propose to group spatial alarms based on their spatial
proximity and check the alarms in selected groups upon
each wakeup. The grouping process proceeds in two
steps. First, all alarms need to be divided into groups
based on spatial proximity with each group associated
with a spatial region. Only when the mobile client
moves into the region marked by a group, the spatial
alarms/subgroups within that group will be checked,
and all other alarms/subgroups belonging to the other
groups are eliminated from alarm checking, leading to
significant saving in terms of computational cost and
energy. For instance, Fig. 4a shows a map of Georgia
Tech with a total of 10 alarms installed on a mobile
client. We group them into three groups with group
one consisting of A1 , A2 , A3 , A4 , group two consisting
of A5 , A6 , A7 , A8 , and group three consisting of A9
and A1 0 (see the three innermost rectangles). All these
three groups together form the fourth group (see the
outer most rectangle in Fig. 4a).
We use the R Tree [13] algorithm to perform the
alarm grouping in a hierarchical fashion as shown in
Fig. 4b. Upon each wakeup the mobile client traverses
down the tree using her current location and terminates
when one of the following conditions becomes true:
(i) a leaf node (alarm) is reached; or (ii) the mobile
client’s location is not bounded by any child’s Minimum
Bounding Rectangle (MBR). The condition (i) signals
that an alarm is satisfied and the appropriate action is
triggered. This approach reduces the number of alarms
to be checked from O(n) in the naive approach of
check-all per wakeup to O(logb n), where n represents
the total number of alarms installed on the mobile
client and b represents the minimum number of alarms
in a group. In our running example shown in Fig. 4a, we
have b = 2 and n = 10.
The R-Tree based alarm grouping algorithm is
effective in terms of energy saving and resource usage
Mobile Netw Appl
Fig. 4 Alarm grouping by
spatial proximity
(a) Alarm Grouping
in general and especially it can handle well, the situations where the mobile client continuously adds new
spatial alarms into the client middleware system as she
moves on the road. However, if the number and the location of the spatial alarms remain unchanged for long
duration of time, we can utilize the Voronoi diagram
to devise a more efficient alarm group algorithm. We
below present two such algorithms, one uses Voronoi
regions and called nearest alarm check algorithm, and
the other uses the Network Voronoi diagrams, called
the road network nearest alarm check algorithm.
4.2 Checking nearest alarm only
The checking nearest alarm only algorithm is suitable
for the scenarios where the number and location of
alarms remain unchanged for long duration of time and
no addition or removal of alarms are issued by the
mobile client. The Nearest Alarm Only optimization
consists of two phases. In the first phase, the two dimensional geographical area of interest is divided into
grid cells of equal size. Then the Voronoi diagram is
overlaid on top of the grid with Voronoi Regions [2, 11]
such that each Voronoi Region has a single nearest
alarm, as shown in Fig. 5a. To facilitate the search for
the nearest alarms for a given mobile client location, we
build a grid cell based dense index, in which each cell
contained in a Voronoi region will point to the spatial
alarm of that region, and each cell that overlaps with
(a) Voronoi Regions
(b) Small grid cells
(b) R Tree of grouped alarms
k Voronoi regions (1 < k < n) will contain k spatial
alarms, each corresponding to one of the k Voronoi
regions.
Given a geographical area of interest, such as the
state of Georgia or the greater area of Atlanta, there
are many mobile clients who will install their personalized spatial alarms. Hence, a third party Voronoi
diagram service provider can be used to generate the
Voronoi diagram for the entire geographical area, and
each mobile client can download the resulting Voronoi
diagram for the area of interest on demand.
In the second phase, upon wakeup the mobile client
uses her current location to locate the grid cell in which
she resides and it takes only O(1) to lookup the nearest
alarm in the case where the grid cell of the client is
contained in a Voronoi region. In the situation where
the mobile client is at boundaries of k Voronoi regions
(1 < k < n), the grid cell in which the client resides
will point to k spatial alarms, all are qualified to be
the ‘nearest’ alarms. In this scenario all the Voronoi
regions overlapped with the current location of the
mobile client need to be considered, and the alarm
check will be performed against the ‘nearest’ alarm
in each of these overlapped Voronoi regions. Clearly,
this approach greatly reduces the time complexity of
lookup the relevant alarms to be checked although it
is only applicable in the specific scenarios where alarms
are not frequently removed or added (since computing
the Voronoi diagram for the entire geographical area
(c) Large grid cells
Fig. 5 Alarm grouping using Voronoi and road network Voronoi approaches
(d) Road Network Voronoi
Mobile Netw Appl
of interest each time when new alarm is added or an
existing alarm is removed can be quite expensive).
However, the storage requirement for such a scheme
is high. Let m denote the total number of grid cells for
the entire grid of the geographical area of interest. We
construct a dense index with m entries. The storage cost
is of the order O(m). With a high value for m, the alarm
lookup time takes O(1) but the storage cost will be at
O(m) (see Fig. 5b). One way to reduce the number m
of grid cells is to increase the grid cell size at the cost of
increased probability of having more than one alarms
qualifying as the nearest alarm for a given grid cell (see
Fig. 5c). This is especially true when the grid cell in
which the mobile client resides overlaps with two or
more Voronoi regions. The larger the size of individual
grid cells is, the larger the number of alarms need to be
checked at each wakeup. At one end of the spectrum,
if there is only one cell covering the entire area of
interest, then this grid cell dense index approach simply
degrades to the naive approach of check-all alarms
upon wakeup. We promote the use of storage to trade
for fast computation since this approach offers better
energy conservation when alarm addition and deletion
are infrequent.
4.3 Checking nearest alarm only with road network
We have discussed the cost of storage for Voronoi region based approach. In order to minimize the storage
requirement, we consider the usage of the Network
Voronoi Diagrams [8, 12] instead without degrading
performance significantly.
The checking nearest alarm with road network algorithm consists of two phases. In the first phase, we need
to build network Voronoi diagram that can partition
the road network among alarm nodes such as each
Voronoi region containing one alarm. Figure 5 illustrates the alarm grouping by dividing the road network
map into network distance based Voronoi regions.
Each alarm (in rectangle shape) has a color associated
with it and the vertices of the road network graph for
which that particular alarm is the nearest (road network
distance) are represented by the same color. Note that
there are certain points which are closer to certain
alarms in terms of Euclidean distance but not so if
Road Network distance is considered. Such alarms in
the Road-Network-Alarm-Grouping strategy would be
grouped with the Road-Network-Nearest-Alarm and
not the Euclidean-Nearest-Alarm.
In the second phase, each of the nodes on the road
network graph has a list of ‘nearest alarms’ based on
road network distance. Thus the alarm checking operation is performed using the O(Nd) algorithm for con-
structing Network Voronoi Diagrams, where N is the
number of nodes (vertices) in the road network graph
and d is the maximum degree of any vertex in the graph.
We give a sketch of the phase one of this algorithm
which performs the road network partitioning among
alarm nodes in Fig. 6 [12].
The road network partitioning among alarm nodes
proceeds in the following steps. First, given a total of n
alarms, and a weighted directed graph G = (V,E) where
V is the number of nodes and E the number of edges,
we first determine the points of intersection of the
alarm area with the roads on the road network graph
and generate a tentative road network graph by adding
these points of intersections as intermediate vertices to
the road network graph. Then we need to partition the
V nodes into n groups (one group for every alarm).
The partitioning process uses three vectors: distance,
f inal and partition and each is of length |V|. At each
iteration, the ith element in the distance vector, denoted
as distancei , represents the currently known shortest
distance to the alarm stored in partitioni . The value in
f inali denotes whether the computations for node i, i.e.,
distancei and partitioni have been finalized or not. If
the node i with the least value of distancei has not been
finalized yet ( f inali = 1), then both the node i and the
nodes which can directly lead up to the node i are chosen. Their distance and partition values are updated if
the new path through node i is of shorter distance. After
all nodes have been exhausted, the values in partition
vector represent the network Voronoi diagram and it
will be used to perform the alarm checking. Figure 5d
shows the partitioning of the vertices of a road network
graph. In the illustration, each alarm has a shape (color)
associated with it and the vertices of the road network
graph for which that particular alarm is the nearest
(road network distance) are represented by the same
shape (color).
Using the road network approach greatly minimizes
storage costs as we can bring down the storage cost
from O(Nd) to O(N) for the N number of nodes in the
road network graph. For instance, N ≈ 35,000 for the
entire city of Atlanta and N ≈ 10,000 for the District of
Columbia [1]. Consider a dire case scenario where each
node on the graph has two ‘nearest’ alarms (requiring
8 bytes of storage with the assumption that each of
the alarms is represented as 32 bit integers), the total
storage cost would be 280 and 80 KB for Atlanta and
DC respectively, which can be easily accommodated in
most of existing mobile devices. At each wakeup, only
one of the two cases will be satisfied: either the mobile
client is on a road (i.e., on the edge of the road network
graph) or the mobile client is at the vertex on the road
network graph. In the former case both the vertices
Mobile Netw Appl
Fig. 6 Constructing the
network Voronoi diagram
at the ends of the road are chosen while in the latter
case the current vertex is chosen. Each of these vertices
will have the list of nearest alarms associated with them
and each of the alarms is checked for satisfaction upon
wakeup. Even in this assumed dire case scenario, it
would require only four (constant) number of alarm
checks on each wakeup.
(a) total energy consumed, (b) total battery lifetime,
(c) alarm density and alarm distribution, and (e) alarm
delivery quality in terms of alarm misses. We below
first describe briefly our simulator design and our experimental setup. Then we present the experimental results, demonstrating the effectiveness of our proposed
middleware architecture for energy-efficient processing
of spatial alarms.
5 Experimental evaluation
5.1 Experimental setup
We have described the four wakeup algorithms and
the three alarm check algorithms. In this section we
evaluate our middleware architecture, the wakeup algorithms, and the alarm check algorithms in terms of
In order to realistically simulate the spatial alarm middleware system on the mobile clients we develop a
simulator that has the environment and work load
generation modules completely separated out from the
Mobile Netw Appl
Fig. 7 Simulator design
specific wakeup and alarm check algorithms used by
the mobile clients. Figure 7 shows a sketch of the key
components of the simulator. The environment module uses the supplied map and generates mobile client
position trace for the duration of time specified. It
also generates the list of alarms based on the alarm
distribution defined by the Alarm Clustering Factor
(see the next subsection for details). The Device module
is independent of the environment module and can be
configured for different mobile clients with different
combinations of the Wake-Up algorithm and the Alarm
Check algorithm. Depending on the settings given by
each mobile client, it plays the client’s position trace file
along with the alarms f ile generated by the environment
module and records the following statistical information: (a) Number of wakeups (b) Number of alarm
checks (c) Number of alarms delivered (d) Storage
cost, apart from the information required for visualizing
the behavior of the system, such as those shown in
Figs. 3 and 8. The Evaluation module comprises scripts
to systematically vary environmental parameters listed
(a) CF: 1
Fig. 8 Clustering factor (CF)
in the simulator parameter table given in Fig. 7, which
generate performance graphs presented in the rest of
Section 5.
In reality the spatial alarms installed by a mobile
client are likely to be restricted in certain regions of
interests, such as a few miles around work place or
home. In order to simulate such behavior we introduce the concept of Clustering Factor. We simulate the
distribution of spatial alarms of a mobile client using
various clustering factors. In general, a clustering factor
of m implies that the alarms are distributed over only
1 th
of the map. Figure 8 depicts the distribution of
m
alarms using three different clustering factors. When
the clustering factor is set to one, it represents the case
where the alarms of a client is distributed over the
entire map. When a clustering factor is set to two, it
represents the scenario where the alarms of a mobile
client are distributed over only one half of the map.
When we set the clustering factor to be a higher value
like ten, it represents the scenario where all the alarms
1 th
are distributed over only 10
of the map.
(b) CF: 3
(c) CF: 10
Mobile Netw Appl
5.2 Estimating energy consumed and battery life
In order to measure the energy consumed and battery lifetime for various wakeup algorithms and alarm
check algorithms, we use the device energy values corresponding to the itsy pocket computer [3] given in
the energy parameter table of Fig. 7 as our reference
model. In the rest of this subsection we show how
one can express the total energy Et consumed as a
function of the number of wakeups (Nw ), the minimum
time duration per wakeup (Tm ), the total time duration
(Tt ), the power consumption of the mobile device while
awake (P100 ), idle (Pi ) and asleep (Ps ), and the Alarm
Check Ratio (Cr ) which represents the ratio of the number of actual alarm checks performed to the maximum
checks that can be performed N A M AX in the minimum
wakeup duration Tm .
First, we compute the total energy Et as the sum of
the energy spent awake and the energy spent in the
sleep state. Let Tw and Ts denote the time periods of
a mobile client in awake and asleep state respectively
and Pw and Ps denote the energy spent per time unit
in awake and asleep state respectively. The following
equation holds.
Et = Tw Pw + Ts Ps
(11)
Let Nw denote the number of wakeups and Tm denote the minimum wakeup duration. We can compute
the time duration for which the mobile client is awake
using the following formula:
Tw = Nw × Tm
(12)
Similarly, the time duration for which the mobile client
is asleep, denoted by Ts , as the difference of total time
duration Tt and the wakeup time duration Tw :
Ts = Tt − Nw Tm
(13)
The energy required for keeping the device (mobile
client) awake is dependent on whether the mobile client
is computing or not. If there are more alarms that the
client has to compute for longer duration of time, then
the power consumption will be higher. After the mobile
client checked for the alarms at each wakeup, it goes to
the idle state for a short duration before going to sleep.
Let Pi denote the power spent by the mobile client during an ideal state, and Cr denote the number of alarm
checks required. The lower the Alarm Check Ratio is,
the lesser the number of alarms need to be checked, and
hence the higher the energy conservation will be. Thus
we can compute the power consumed during wakeup,
denoted as Pw , by the following formula:
Pw = Cr P100 + (1 − Cr )Pi
(14)
Where Cr can be defined in terms of the number of
actual alarm checks (NC ), the maximum number of
alarm checks that can be performed (N A M AX ), and the
number of wakeups (Nw ):
Cr =
NC
N A M AX ∗ Nw
(15)
By substituting Eqs. 12, 13, and 14 in Eq. 11, we
obtain the Total Energy consumed as follows:
Et = Nw Tm (Cr P100 + (1 − Cr )Pi ) + (Tt − Nw Tm )Ps
(16)
Let Cb be the battery capacity, Tt be the total time of
the experiment, and Et represent the energy spent during Tt . The battery lifetime of a mobile client, denoted
by Tb , can be calculated by the following equation:
Tb =
Cb × T t
Et
(17)
5.3 Experimental results
In this section we present a set of experimental results.
Our results demonstrate three important conclusions.
First, the proposed wakeup and alarm check algorithms
offer significant (up to 6.4 times) reduction in terms
of energy consumption in comparison to the naive
approach with periodic wakeups followed by checking
all alarms per wakeup. Second, the Safe Road Distance
with Max Speed wakeup algorithm offers the maximum
energy conservation with 100% alarm delivery guarantee. Third but not the least, the alarm grouping check
algorithm is the most flexible among the alternative
alarm check strategies and offers significant reduction
in terms of energy consumption, and significant improvement (up to 50%) in battery life when the number
of alarms is high. The Nearest-only Alarm and the Network Nearest-only Alarm algorithms offer even better
improvements in terms of energy consumption but have
limited applicability and can only be used in those cases
where addition and removal of alarms are less frequent
and low in numbers.
5.3.1 Overall energy consumption and battery lifetime
Figures 9, 10, 11 compare energy and battery life time
for a mobile device by varying the number of alarms
per map tile for different wakeup algorithms (Fig. 9),
different alarm check algorithms (Fig. 10), and five alternative combinations of wakeup algorithm and alarm
check algorithm (Fig. 11). We used the measurements
for the Itsy Pocket Computer [3] given in the energy
Mobile Netw Appl
Fig. 9 Performance of WakeUp strategies
Fig. 10 Performance of check strategies
Fig. 11 Performance of WakeUp-Check strategy combinations
Mobile Netw Appl
parameter table of Fig. 7 for these experiments. The
results were plotted using the equations in Section 5.2.
Apart from the values listed in the table in Fig. 7,
we want to demonstrate how the algorithms perform
when the number of alarms reaches high values and
assume that the mobile client can check only 100 alarms
(T A M AX ) in 0.001 h. Though in reality this number might
be higher, by making such a pessimistic assumption we
get to stress-test the system and see its behavior at the
boundary conditions.
Figure 9 (left) compares the energy consumption of
the individual wakeup algorithms by fixing an alarm
check algorithm and we use the naive alarm check algorithm in this case. It shows that our wakeup algorithms
(especially the Safe Road Distance algorithms) perform
close to six times better than the naive algorithm of
periodic wakeup and check all alarms per wakeup. The
increase in energy consumption is almost linear with
the increase in the number of alarms for all four of
our wakeup algorithms because we use the naive alarm
checking strategy in all the cases. Figure 9 (right) shows
the corresponding lifetime of the battery for each of
the wakeup algorithms. As shown in the figure, all our
wakeup algorithms offer significant improvement in
terms of the battery lifetime of the mobile client and the
battery lifetime converges when the number of alarms
are high, indicating that by using the wakeup algorithms
alone, it is not always sufficient to reduce the energy
consumption involved. For such scenarios, alarm check
algorithms are necessary for energy conservation purpose.
Figure 10 (left) compares the energy consumption of
the alarm check algorithms by fixing the wakeup Strategy (Periodic WakeUp Strategy is used in this case). We
show that all three proposed alarm check strategies turn
out to outperform the naive approach that checks all
alarms upon each wakeup, especially when the number
of alarms are high. The Energy consumption with the
naive check algorithm grows almost linearly while with
our proposed alarm check algorithms, the increase in
terms of energy consumption is much slower. This is
to be expected because the naive approach checks all
alarms upon each wakeup, which can be significantly
degraded when the number of alarms is high. Figure 10
(right) shows the corresponding effect on the battery
lifetime. The flexible alarm grouping based checking
algorithm offers close to 86% improvement over the
conventional approach. Further, the trend shows that
these improvements grow with the increase in the number of alarms.
Finally, we examine the effectiveness of the wakeup
and alarm check combinations. Figure 11 shows some
of the interesting combinations. Due to the space con-
straint, we omit for a large number of possible combinations. The battery lifetime plot in Fig. 11 (right)
is similar to Fig. 9 (right). Even at higher number of
alarms per map tile, we show that the battery lifetime
of a mobile client is considerably better in the wakeup
case, reflecting the need for a good alarm checking
strategy that parts from the wakeup algorithms.
5.3.2 Ef fect of changes in number of alarms
Figure 12 (left) compares the variation of Wake-Up
frequencies with the increase in number of alarms. As
one would expect the Periodic Wake-Up strategy is
indifferent with the increase in the number of alarms.
When the mobile client has only one alarm per map tile,
the proposed wakeup algorithms reduce the number
of wakeups by almost an order of magnitude. The two
Safe Road Distance algorithms perform better than the
Safe Distance algorithms. The wakeup algorithms that
use expected speed reduce the frequency of wakeups
further. However, as shown in Fig. 14 the expected
speed algorithms may not guarantee the delivery of all
alarms at all times. Also all the proposed wakeup algorithms increase the wakeup frequency as the number of
alarms increases and similarly, all the check algorithms
increase the number of alarm checks as the number of
alarms increases. This is expected because with increase
in number of alarms the average minimum distance to
any alarm would be reduced and hence both the Safe
Distance and Safe Road Distance Wake-Up algorithms
will allow the mobile client to sleep only for shorter
durations. Note that in this set of experiments we use
a clustering factor of two and all the alarms are limited
to one half of the map, resulting in the frequency of
wakeups to level out for high number of alarms.
Figure 12 (right) compares the Alarm Check algorithms as the number of alarms varies. As one would
expect, with increase in number of alarms, the number
of checks per wakeup also increases. Such an increase
is linear for the naive approach that checks against all
the alarms at each wakeup. However, for the alarm
grouping approach, which uses a R-Tree [13] based
lookup, the increase is sub-linear. The Nearest Alarm
Only and the Nearest Road Network Distance Alarm
Only offer almost constant number of checks at each
wakeup irrespective of the number of the alarms installed. In reality, with increase in number of alarms,
the number of checks for these two strategies can also
go up but is almost negligible in practical scenarios.
Only in some extreme cases where almost all the alarms
are overlapping with each other, these strategies would
end up requiring checks on all the alarms at each
wakeup.
Mobile Netw Appl
Fig. 12 Effect of changes in number of alarms
5.3.3 Ef fect of changes in alarm distribution
Figure 13 shows the performance comparison of periodic wakeup with the four proposed wakeup algorithms
with an increase in the Clustering Factor. We see that
as the Clustering factor increases, the Wake-Up algorithms tend to perform better. This is simply because
when the clustering factor is high, it implies that the
alarms are clustered around a portion of the map, thus
the minimum distance to the nearest alarm from the
client’s current location is higher, and the client has
longer time to sleep.
Figure 13 (right) shows the effect of changing clustering factor on the Alarm Check Strategies. It is interesting to note that with increase in clustering factor,
the grouping near by alarms algorithm performs particularly well in terms of performance improvement.
This is because the closer the alarms are distributed,
Fig. 13 Effect of clustering of alarms
the smaller the size of the Minimum Bounding Rectangle (MBR) required to enclose them, and hence the
lesser the number of alarm checks to be performed
at each wakeup. It is also important to note that at
higher clustering factors there is a slight increase in the
number of alarm checks for the Nearest Alarm Only
algorithm. This is because when alarms get very close to
each other, grid cells used by this strategy tend to have
more than one ‘Nearest’ alarms which in turn result in
checking of more than one xfaction.
5.3.4 Alarm delivery quality
Figure 14 evaluates the quality of the alarm delivery provided by each of the wakeup algorithms with
variations in the Number of Alarms and the Clustering Factor respectively. Periodic, Safe Distance with
Mobile Netw Appl
Fig. 14 Alarm delivery quality
Maximum Speed and Safe Road Distance with Maximum Speed perform perfectly in terms of 100% delivery
of alarms. However, Safe Road Distance with Expected
Speed misses a few of alarms. Note that in some cases
only about 95% of the alarms are delivered. One would
expect the same behavior for the Safe Distance with Expected Speed approach. But surprisingly in most of the
cases the Safe Distance with Expected Speed approach
manages to have a 100% alarm recall rate. We believe
that this is because the use of Safe Euclidean Distance
played a critical role in offsetting the errors produced
in expected speed estimation. This experiment shows
that the Expected Speed algorithms perform better in
terms of reducing the number of wakeups and saving
energy. However they are not flawless in delivering
alarms with 100% recall, and hence should be used
with care. It is important to note that the alarm check
algorithms do not affect the alarm delivery quality.
Due to the space constraint we omit the corresponding
experimental results in this paper.
5.3.5 Ef fect of alarm group size
Figure 15 shows the effect of the alarm group sizes
on the final computation required for spatial alarm
processing. A maximum alarm group size of m implies
that the maximum fan out of the R Tree [13] used to
group the alarms is m. In other words, no more than
m alarms can be put together to form a group and no
more than m of such groups can be put together into
a higher level group, and so forth. It is interesting to
note that the larger the group size value of m is, the
higher the number of alarms will need to be checked,
and hence the more worse the performance of alarm
process will be. This is particularly the case when the
alarms are scattered all around the map (CF = 1).
5.3.6 Storage requirements for nearest alarm only
check algorithm
Fig. 15 Effect of alarm group size
We measure the tradeoffs between the Storage requirement vs minimizing number of checks for the Nearest Alarm Only Alarm Checking algorithm. In order
to make the effects more pronounced we used forty
alarms (on the same map tile of about 1.61km2 area)
with a clustering factor of three such that a large number of alarms would be clustered around a small region.
Our experimental results show that with increase in
number of grid cells per row of the map tile, the storage
requirements for the dense index also increase rapidly
(Fig. 16 (left)), while at the same time the number of
checks required at each wakeup is reduced at a very
high rate (Fig. 16 (right)).
Mobile Netw Appl
Fig. 16 Storage requirements
6 Related work and conclusion
The idea of location based reminder and its use have
been discussed in several human–computer interaction
projects in the recent years [4, 6, 15, 16, 20, 21]. These
papers primarily concentrate on usability of such location reminder systems from the human–computer
interaction point of view. For instance, [14] focused
on identifying the user’s logical location by mapping
information such as IP Address, latitude and longitude
information in GPS readings. The logical location could
then be used to automatically connect the mobile client
to local resources such as printers.
On the other hand, there are much works like [9, 10,
18], which describe general strategies to minimize energy consumptions on mobile devices, such as Dynamic
voltage scaling, and to adapt mobile applications to
consume lesser energy. There has also been research in
the area of continuous spatial queries like in [17, 19, 22].
Our work in this paper distinguishes itself primarily
along two dimensions—(a) addressing the special class
of one time queries i.e. detecting the spatial validity of
the alarm and triggering the action if conditions are
satisfied as opposed to continuously keeping track of
the objects that satisfy the nearest-neighbor or range
query and (b) by addressing the energy consumption
aspects as opposed to addressing only scalability or
performance efficiency.
We have presented a middleware architecture for
energy efficient processing of spatial alarms on mobile
clients. Our approach to spatial alarms provides two
systematic methods for minimizing energy consumption on mobile clients. First, we introduce the concept
of safe period to reduce the number of unnecessary
wakeups for spatial alarm evaluation, enabling mobile clients to sleep for longer periods of time. We
show that our safe period techniques can significantly
minimize the energy consumption on mobile clients
compared to periodic wakeups while maintaining the
desired accuracy and timeliness of spatial alarms. Second, we develop a suite of techniques for minimizing
the number of location triggers to be checked for spatial
alarm evaluation at each wakeup. This further reduces
the computation cost and energy expenditure on mobile clients. We evaluate the scalability and energyefficiency of our wakeup and alarm check approach
using a road network simulator. The experiments show
that our spatial alarms middleware architecture offers
significant improvements on battery lifetime of mobile
clients, while maintaining high quality of spatial alarm
services compared to the conventional approach of
periodic wakeup and checking all alarms upon each
wakeup.
Acknowledgements This work is partially supported by grants
from NSF CISE under CNS, NetSE, and CyberTrust, a grant
from AFOSR, an IBM SUR grant, and an IBM faculty award,
and an Intel research council grant. We would also like to thank
Ramesh Chandran, Kipp Jones, Calton Pu, and the mobile computing group within the DiSL lab for the valuable discussions and
help with the experiments.
References
1. US Census Bureau, Tiger, tiger/line and tiger-related products. Available at http://www.census.gov/geo/www/tiger/
2. Aurenhammer F (1991) Voronoi diagrams a survey of a
fundamental geometric data structure. ACM Comput Surv
23(3):345–405
3. Bartlett JF, Brakmo LS, Farkas KI, Hamburgen WR, Mann
T, Viredaz MA, Waldspurger CA, Wallach DA (2000) The
itsy pocket computer. Technical Report 2000/6, COMPAQ
Western Research Laboratory
Mobile Netw Appl
4. Bazinette V, Cohen NH, Ebling MR, Hunt GDH, Lei H,
Purakayastha A, Stewart G, Wong L, Yeh DL (2001) An
intelligent notification system. IBM research report
5. Cherkassky BV, Goldberg AV, Radzik T (1993) Shortest
paths algorithms: theory and experimental evaluation. Math
Program 73:129–174
6. Dey AK, Abowd GD (2000) Cybreminder: a context-aware
system for supporting reminders. In: Handheld and ubiquitous computing: second international symposium, HUC 2000,
Bristol, UK proceedings
7. Dijkstra EW (1959) A note on two problems in connection
with graphs. Numerische Math 1:269–271
8. Erwig M (2000) The graph voronoi diagram with applications. In: Networks, pp 156–163
9. Flautner K, Mudge T (2002) Vertigo: automatic
performance-setting for linux. SIGOPS Oper Syst Rev
36(SI):105–116
10. Flinn J, Satyanarayanan M (1999) Energy-aware adaptation
for mobile applications. In: SOSP ’99: proceedings of the seventeenth ACM symposium on operating systems principles.
ACM Press, New York, NY, USA, pp 48–63
11. Fortune S (1997) Voronoi diagrams and delaunay triangulations. In: Goodman JE and OŠRourke J (eds) Handbook
of discrete and computational geometry. CRC, Boca Raton,
pp 377–388
12. Graf M, Winter S (2003) Netzwerk-voronoi-diagramme.
Österr Z Vermess Geoinf 91:166–174
13. Guttman A (1984) R-trees: a dynamic index structure for
spatial searching. In: SIGMOD ’84: proceedings of the 1984
ACM SIGMOD international conference on management of
data. ACM, New York, pp 47–57
14. Heidemann J, Shah D (2000) Location-aware scheduling with
minimal infrastructure. In: USENIX conference proceedings.
USC/Information Sciences Institute, USENIX, San Diego,
pp 131–138
15. Kim SW, Kim MC, Park SH, Jin YK, Choi WS (2004) Gate
reminder: a design case of a smart reminder. In: DIS ’04:
Proceedings of the 2004 conference on designing interactive
systems. ACM, New York, pp 81–90
16. Ludford PJ, Frankowski D, Reily K, Wilms K, Terveen L
(2006) Because i carry my cell phone anyway: functional
location-based reminder applications. In: CHI ’06: proceedings of the SIGCHI conference on human factors i computing
systems. ACM, New York, pp 889–898
17. Mokbel MF, Xiong X, Aref WG (2004) Sina: scalable incremental processing of continuous queries in spatio-temporal
databases. In: In SIGMOD, pp 623–634
18. Mudge T (2001) Power: a first-class architectural design
constraint. In: Computer. Michigan Univ., Ann Arbor,
pp 52–58
19. Prabhakar S, Xia Y, Kalashnikov DV, Aref WG, Hambrusch
SE (2002) Query indexing and velocity constrained indexing:
scalable techniques for continuous queries on moving objects.
IEEE Trans Comput 51(10):1124–1140
20. Sadeh NM, Gandon FL, Kwon OB (2005) Ambient intelligence: the mycampus experience. In: CMU-ISRI-05-123
21. Sohn T, Li KA, Lee G, Smith I, Scott J, Griswold WG
(2005) Place-its: a study of location-based reminders on mobile phones. In: UbiComp 2005: ubiquitous computing
22. Zhang J, Zhu M, Papadias D, Tao Y, Lee DL (2003)
Location-based spatial queries. In: In SIGMOD, pp 443–
454